The Pandemonium Architecture

In 1958, before there were neural networks, before there was machine learning as a discipline, Oliver Selfridge described a machine made of competing demons and called it Pandemonium—after the capital of Hell in Milton’s Paradise Lost, the place whose Greek name means “all the demons.” The architecture has four levels: data demons holding raw input; feature demons, each watching for one specific fragment of the signal and shrieking its confidence; cognitive demons that combine features into candidate interpretations; and a decision demon that listens for the loudest cognitive demon and declares the answer. No single demon understands anything. The understanding—such as it is—is the emergent property of the shouting match. This was, at the time, a heresy against the serial, rule-following, symbol-manipulating picture of thought. It turned out to be an accurate description of how every neural network on Earth actually works. The “deep” in deep learning is nothing other than the depth of Selfridge’s demon hierarchy, multiplied. The training of connection weights in response